-
Notifications
You must be signed in to change notification settings - Fork 1
/
cnn_lstm.py
134 lines (115 loc) · 5.03 KB
/
cnn_lstm.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
# -*- coding: utf-8 -*-
"""
This is an attempt to use a single word embedding for multiple filters by combining
the Graph and Sequential containers.
@author: ameasure
"""
import datetime
import pandas as pd
import numpy as np
np.random.seed(42)
from sklearn.preprocessing import LabelEncoder
import theano
import keras
from keras.models import Sequential, Graph
from keras.layers.convolutional import Convolution1D, MaxPooling1D, Convolution2D, MaxPooling2D
from keras.layers.core import Dense, Dropout, Activation, Flatten, Reshape, Merge
from keras.layers.recurrent import LSTM
from keras.layers.normalization import BatchNormalization
from keras.layers.embeddings import Embedding
from keras.optimizers import SGD
from keras.utils import np_utils
from keras.preprocessing.text import Tokenizer
from keras.preprocessing.sequence import pad_sequences
from keras.regularizers import l2
from embed_preprocessing import get_initial_embeddings
from msha_extractor import get_data
#output needs to be (nb_samples, timesteps, input_dim)
def generate_seq_to_one_hot(X, Y, vocab_size, batch_size):
n_samples = len(X)
seq_len = len(X[0])
start = 0
while 1:
stop = start + batch_size
chunk = X[start: stop]
slices = []
for i, seq_indexes in enumerate(chunk):
x_slice = np.zeros([seq_len, vocab_size])
x_slice[np.arange(seq_len), seq_indexes] = 1
slices.append(x_slice)
x_out = np.stack(slices, axis=0)
y_out = Y[start: stop]
start += batch_size
if (start + batch_size) > n_samples:
print 'reshuffling, %s + %s > %s' % (start, batch_size, n_samples)
remaining_X = X[start: start + batch_size]
remaining_Y = Y[start: start + batch_size]
random_index = np.random.permutation(n_samples)
X = np.vstack((remaining_X, X[random_index, :]))
Y = np.vstack((remaining_Y, Y[random_index, :]))
start = 0
n_samples = len(X)
yield x_out, y_out
max_len = 100
#embedding_size = 300
#batch_size = 128
gen_batch_size = 10
code_type = 'ACTIVITY_CD'
raw_train, raw_valid, raw_test = get_data(n_train=47500, n_valid=2500, n_test=10000)
raw_train_labels = raw_train[code_type]
raw_valid_labels = raw_valid[code_type]
raw_test_labels = raw_test[code_type]
labeler = LabelEncoder()
labels = set(raw_train[code_type].tolist() + raw_test[code_type].tolist())
labeler.fit(list(labels))
nb_classes = len(set(labels))
print('nb_classes = %s' % nb_classes)
y_train = labeler.transform(raw_train_labels)
Y_train = np_utils.to_categorical(y_train, nb_classes)
y_valid = labeler.transform(raw_valid_labels)
Y_valid = np_utils.to_categorical(y_valid, nb_classes)
y_test = labeler.transform(raw_test_labels)
Y_test = np_utils.to_categorical(y_test, nb_classes)
print 'Tokenizing X_train'
tokenizer = Tokenizer(nb_words=10000)
tokenizer.fit_on_texts(raw_train['NARRATIVE'])
tokenizer.word_index['BLANK_EMBEDDING'] = 0
X_train = tokenizer.texts_to_sequences(raw_train['NARRATIVE'])
X_valid = tokenizer.texts_to_sequences(raw_valid['NARRATIVE'])
X_test = tokenizer.texts_to_sequences(raw_test['NARRATIVE'])
X_train = pad_sequences(X_train, maxlen=max_len)
X_valid = pad_sequences(X_valid, maxlen=max_len)
X_test = pad_sequences(X_test, maxlen=max_len)
print('X_train shape:', X_train.shape)
print('X_valid shape:', X_valid.shape)
print('X_test shape:', X_test.shape)
#initial_embeddings = get_initial_embeddings(tokenizer.word_index)
vocab_size = len(tokenizer.word_index)
#filter_length = 7
#nb_filter = 300
model = Sequential()
#model.add(Embedding(vocab_size, embedding_size, input_length=max_len, weights=[initial_embeddings]))
#model.add(Convolution1D(nb_filter=nb_filter,
# filter_length=filter_length,
# border_mode='valid',
# activation='relu',
# subsample_length=1))
model.add(LSTM(100, dropout_W=0.5, dropout_U=0.5, return_sequences=True, input_shape=(max_len, vocab_size)))
model.add(MaxPooling1D(pool_length=max_len))
model.add(Flatten())
model.add(Dropout(0.5))
model.add(Dense(output_dim=nb_classes))
model.add(Activation('softmax'))
model.compile(loss='categorical_crossentropy',
optimizer='adam')
#model.fit(X_train, Y_train, batch_size=batch_size, nb_epoch=15,
# verbose=True, show_accuracy=True, validation_split=.05)
train_generator = generate_seq_to_one_hot(X_train, Y_train, vocab_size, batch_size=gen_batch_size)
valid_generator = generate_seq_to_one_hot(X_valid, Y_valid, vocab_size, batch_size=gen_batch_size)
model.fit_generator(generator=train_generator, samples_per_epoch=len(X_train), nb_epoch=30,
show_accuracy=True, validation_data=valid_generator, nb_val_samples=len(X_valid))
test_generator = generate_seq_to_one_hot(X_test, Y_test, vocab_size, batch_size=gen_batch_size)
score = model.evaluate_generator(generator=test_generator, val_samples=len(X_test),
show_accuracy=True)
print('Test score:', score[0])
print('Test accuracy:', score[1])